Proceedings of the Tenth Workshop on Innovative Use of NLP for

Transcription

Proceedings of the Tenth Workshop on Innovative Use of NLP for
Judging the Quality of Automatically Generated Gap-fill Question using
Active Learning
Nobal B. Niraula and Vasile Rus
Department of Computer Science and Institute for Intelligent Systems
The University of Memphis
Memphis, TN, 38152, USA
{nbnraula,vrus}@memphis.edu
Abstract
online education platforms. These systems typically
deliver knowledge to learners via video streaming
or direct interaction with the system, e.g. dialogue
based interaction. If adaptive to individual learners,
such online platforms for learning must assess learners’ knowledge before, during, and after students’
interaction with the platform. For instance, in order to identify knowledge deficits before and/or after a session a pre- and/or post-test can be used. The
knowledge deficits discovered based on the pre-test
can guide the online platform to select appropriate
instructional tasks for the learner. Furthermore, the
pre- and post-test can be used to measure the learning gains with the online platform, e.g. by subtractic the pre-test score from the post-test score. The
bottom line is that assessment is critical for adaptive instruction. Various kinds of questions are used
to assess students’ knowledge levels varying from
True/False questions to multiple choice questions to
open answer questions.
In this paper, we propose to use active learning for training classifiers to judge the quality of gap-fill questions. Gap-fill questions are
widely used for assessments in education contexts because they can be graded automatically
while offering reliable assessment of learners’
knowledge level if appropriately calibrated.
Active learning is a machine learning framework which is typically used when unlabeled
data is abundant but manual annotation is slow
and expensive. This is the case in many Natural Language Processing tasks, including automated question generation, which is our focus. A key task in automated question generation is judging the quality of the generated
questions. Classifiers can be built to address
this task which typically are trained on human
labeled data. Our evaluation results suggest
that the use of active learning leads to accurate
classifiers for judging the quality of gap-fill
questions while keeping the annotation costs
in check. We are not aware of any previous effort that uses active learning for question evaluation.
1
Introduction
Recent explosion of massive open online courses
(MOOCs) such as Coursera1 and Udacity2 and the
success of Intelligent Tutoring Systems (ITSs), e.g.
AutoTutor (Graesser et al., 2004) and DeepTutor
(Rus et al., 2013), at inducing learning gains comparable to human tutors indicate great opportunities for
1
2
http://www.coursera.org
http://www.udacity.com
Indeed, a main challenge in online learning platforms such as MOOCs and ITSs is test construction (assessment question generation). Automated
test construction is a demanding task requiring significant resources. Any level of automation in
question generation would therefore be very useful for this expensive and time-consuming process.
In fact, it has been proven that computer-assisted
test construction can dramatically reduce costs associated with test construction activities (Pollock
et al., 2000). Besides test construction, automatic
question generation are very useful in several other
applications such as reading comprehension (Eason et al., 2012), vocabulary assessment (Brown et
196
Proceedings of the Tenth Workshop on Innovative Use of NLP for Building Educational Applications, 2015, pages 196–206,
c
Denver, Colorado, June 4, 2015. 2015
Association for Computational Linguistics
al., 2005), and academic writing(Liu et al., 2012).
Consequently, particular attention has been paid
by Natural Language Processing (NLP) and educational researchers to automatically generating several types of questions. Some examples include multiple choice questions (Mitkov et al., 2006; Niraula
et al., 2014), gap-fill questions (Becker et al., 2012)
and free-response questions (Mazidi and Nielsen,
2014a; Heilman and Smith, 2009). The more general problem of question generation has been systematically addressed via shared tasks (Rus et al.,
2010).
Mitkov et al. (2006) reported that automatic question construction followed by manual correction is
more time-efficient than manual construction of the
questions alone. Automated method for judging the
question quality would therefore make the question
generation process much more efficient. To this end,
we present in this paper an efficient method to rank
gap-fill questions, a key step in generating the questions. We formulate the problem next.
1.1
Gap-fill Question Generation
Gap-fill questions are fill-in-the-blank questions
consisting of a sentence/paragraph with one or
more gaps (blanks). A typical gap-fill question is
presented below:
Newton’s
law is relevant after the mover
doubles his force as we just established that there is
a non-zero net force acting on the desk then.
The gap-fill question presented above has a word
missing (i.e. a gap). A gap-fill question can have one
more than one gaps too. Students (test takers) are
supposed to predict the missing word(s) in their answer(s). Gap-fill questions can be of two types: with
alternative options (key and distractors) and without
choices. The former are called cloze questions and
the latter are called open-cloze questions. In this paper, we use the term gap-fill question as an alternative to open-cloze question.
The attractiveness of gap-fill questions is that they
are well-suited for automatic grading because the
correct answer is simply the original word/phrase
corresponding to the gap in the original sentence.
As a result they are frequently used in educational
contexts such as ITSs and MOOCs.
197
Figure 1: A pipeline for gap-fill question generation
A typical pipeline to automatically generate gapfill questions is shown in Figure 1. It follows the
three steps paradigm for question generation (Rus
and Graesser, 2009) : Sentence Selection, Candidate
Generation (overgeneration) and Candidate Selection (ranking).
Step 1 - Sentence Selection: To generate gap-fill
questions, a set of meaningful sentences are needed
first. The sentences can be selected from a larger
source, e.g. a chapter in a textbook, using particular
instructional criteria such as being difficult to comprehend or more general informationl criteria such
as being a good summary of the source (Mihalcea,
2004) or directly from subject matter experts.
Step 2 - Candidate Generation: This step generates a list of candidate questions (overgeneration)
from the target sentences selected in Step 1. The
simplest method might be a brute force approach
which generates candidate questions by considering
each word (or a phrase) as a gap. A more advanced
technique may target the content words as gaps or
exploit the arguments of semantic roles for the gaps
(Becker et al., 2012). An example of overgeneration
of questions is shown in Table 1.
Step 3 - Candidate selection: Not all of the questions generated in the candidate generation step are
of the same quality. The classes can be Good, Okay
and Bad as in Becker et al. (2012) or simply the binary classes Good and Bad. Good questions are the
questions that ask about key concepts from the sentence and are reasonable to answer, Okay questions
are questions that target the key concepts but are
difficult to answer (e.g. too long, ambiguous), and
Bad questions are questions which ask about unimportant aspect of the sentence or their answers are
easy to guess from the context. The candidate selection step is about rating the question candidates. Supervised machine learning models are typically em-
Bad
Good
Good
Good
Bad
Okay
Bad
Good
Okay
Bad
........ net force is equal to the mass times its acceleration.
The ........ force is equal to the mass times its acceleration.
The net ........ is equal to the mass times its acceleration.
........ is equal to the mass times its acceleration.
The net force ........ equal to the mass times its acceleration.
The net force is ........ to the mass times its acceleration.
The net force is equal ........ the mass times its acceleration.
The net force is equal to the ........ times its acceleration.
The net force is equal to the mass ........ its acceleration.
The net force is equal to the mass times ........ acceleration.
Table 1: Typical overgenerated questions from a sentence with their ratings Good, Okay and Bad.
ployed in the form of classifiers to label the candidate questions as Good, Okay, or Bad.
1.2
Question Quality
Question quality can be judged linguistically or pedagogically. In linguistic evaluation, questions are
evaluated with respect to whether they are grammatically and semantically correct. In pedagogical evaluation, questions are evaluated to see whether they
are helpful for understanding and learning the target
concepts. Our focus here is on the pedagogical evaluation of automatically generated gap-fill questions
since they are always linguistically correct.
The third step i.e. candidate selection is expensive when supervised approaches are used because
model building in supervised learning requires large
amount of human annotated examples. The advantage of supervised methods, however, is that their
performances are in general better than, for instance,
that of unsupervised methods. As such, ideally,
we would like to keep the advantages of supervised
methods while reducing the costs of annotating data.
Such a method that offers a good compromise between annotation costs and performance is active
learning, which we adopt in this work. Such models
are always attractive choices especially when there
is a limited budget e.g. fixed annotation time / cost,
a highly probable case.
Active learning and interactive learning are two
well-known approaches that maximize performance
of machine learning methods for a given budget.
They are successfully applied for rapidly scaling dialog systems (Williams et al., 2015), parts-of-speech
tagging (Ringger et al., 2007), sequence labeling
198
(Settles and Craven, 2008), word sense disambiguation (Chen et al., 2006), named entity tagging (Shen
et al., 2004), etc. Instead of selecting and presenting to an annotator a random sample of unlabeled
instances to annotate, these approaches intelligently
rank the set of unlabeled instances using certain criteria (see Section 3) and present to the annotator
the best candidate(s). This characteristic of active
learning and interactive labeling hopefully demands
fewer instances than random sampling to obtain the
same level of performance.
In this paper, we propose an active learning based
approach to judge the quality of gap-fill questions
with the goal of reducing the annotation costs. We
are not aware of any previous effort that uses active
learning for question generation. We chose active
learning particularly because it is well-suited when
unlabeled data is abundant but manual annotation is
tedious and expensive. As mentioned, this is the
case in gap-fill question question generation in overgeneration approaches when plenty of questions are
available but their quality needs to be specified. The
remaining challenge is to judge the quality of these
questions. Our plan is to build a probabilistic classifier at reduced costs that would automatically label
each candidate questions as good or bad using an
active learnign approach.
The rest of the paper is organized as follows. In
Section 2, we present the relevant works. In Section 3, we present the active learning techniques that
we are going to employ. In Section 4 and Section
5, we describe our experiments and results respectively. We present the conclusions in Section 6.
2
Related Works
Currently, statistical and machine learning based approaches are the most popular approaches that are
used to rank the automatically generated questions
of various kinds e.g. free-response (e.g. What,
When etc.) and gap-fill questions. For example,
Heilman et al. (2010) used logistic regression, a
supervised method, to predict the acceptability of
each free-response question candidate. The candidate questions were automatically generated by using a set of rules. They used fifteen native Englishspeaking university students for the construction of
training examples required for building the logistic
regression model.
Hoshino and Nakagawa (2005) proposed a machine learning approach to generate multiple-choice
questions for language testing. They formed a question sentence by deciding the position of the gap i.e.
missing word(s). To decide whether a given word
can be left blank (i.e. serve as a gap) in the declarative stem, they trained classifiers using the training
instances which were generated by collecting fillin-the-blank questions from a TOEIC preparation
book. The positive examples were the exact blank
positions in the question from the book whereas the
negative examples were generated by shifting the
blank position.
Similarly, Becker et al. (2012) proposed Mind the
Gap system that applied logistic regression to rank
automatically generated gap-fill questions. They
used text summarization technique to select useful
sentences from text articles for which gap-fill questions are to be generated. From each of the selected
sentence, it generated potential gap-fill candidates
using semantic constraints. Each candidate question was then labeled by four Amazon’s Mechanical Turkers to one of Good, Bad and Okay classes.
In total, 85 unique Turkers were involved in the annotation. The data set was used to build a logistic
regression classifier and ranked the candidate questions. They reported that the classifier largely agreed
with the human judgment on question quality.
In recent works Mazidi and Nielsen (2014a;
2014b) generated free-response questions from sentences by using the patterns which were manually authored by exploiting the semantic role labels.
They evaluated the questions linguistically and ped199
agogically using human annotators and reported that
their systems produced higher quality questions than
comparable systems. The main limitation of their
approaches is that they do not exploit the examples
obtained from the annotation process to evaluate unseen (or not yet evaluated) questions. Moreover,
their approaches do not provide any ranking for the
questions they generated using those patterns.
3
Active Learning for Judging Question
Quality
As mentioned before, active learning fits well when
abundant data can be available but manual labeling
costs are high. As a result, the technique has been
applied to many NLP tasks such as text classification, Word Sense Disambiguation, sequence labeling, and parsing. We use active learning for guiding
our annotation process for judging the quality of automatically generated gap-fill questions.
3.1
Active Learning Algorithms
An active learning system mainly consists of a classification model and querying algorithm. Typically
the classification models are the probabilistic classifiers such as Na¨ıve Bayes and Logistic Regression
which provide a class probability distribution for a
given instance. Querying algorithms/functions actively choose unlabeled instance samples by exploiting these probabilities.
We follow the standard pool-based active learning
algorithm as shown in Algorithm 1. It starts with a
set of initially labeled instances (seed examples) and
a set of unlabeled instances (U ). A new model is
built using the labeled examples in L. Next, a batch
of instances are extracted from the unlabeled set U
using a query function f (.) and then the selected instances are labeled by human judges. The new labeled instances are added to the labeled list L. The
process repeats until a stopping criterion is met. The
criteria could be the number of examples labeled,
expected accuracy of the model, or something else.
3.1.1
Querying Algorithms
Many query functions exist. They differ on how
they utilize the class probability distributions. We
use two variants of query functions: uncertainty
sampling and query by committee sampling.
Input: Labeled instances L, unlabeled
instances U , query batch size B, query function
f (.) ;
while some stopping criterion do
θ = Train the model using L;
for i = 1 to B do
b∗i = arg maxu∈U f (u);
L = L ∪ < b∗i , label(b∗i ) >;
U = U − b∗i ;
end
end
Algorithm 1: Pool-based active learning algorithm
A. Query By Uncertainty or Uncertainty Sampling
Uncertainty sampling chooses the samples for
which the model’s predictions are least certain.
These examples reside very near to the decision
boundary. We use three functions that predict the
samples in the decision boundary.
(a) Least Confidence: This function chooses the
sample x that has the highest fLC (.) score and is
defined as : fLC (x) = 1 − P (y ∗ |x; θ) where y ∗ is
the most likely class predicted by the model (Settles
and Craven, 2008).
(b) Minimum Margin: This function chooses the
sample x that has the least fM M (.) score and is
defined as: fM M (x) = |P (y1∗ |x; θ) − P (y2∗ |x; θ)|
where y1∗ and y2∗ are the first and the second most
likely classes predicted by the model (Chen et al.,
2006).
(c) Entropy: This function chooses the sample x
that has the highest entropy i.e. fEN (.) score and
P
is defined as: fEN (x) = − C
c=1 P (yc |x; θ) ∗
log(P (yc |x; θ)) where C is the total number of
classes (Chen et al., 2006).
B. Query By Committee
Our query by committee sampling algorithm consists of a committee of models. These models are
trained on the same labeled examples but learn different hypotheses. We compute for a given instance
the class distribution mean over all committee members and assume that the mean scores represent the
votes received from the committee. Next we apply
fLC (.) , fM M (.) and fEN (.) over the mean class
200
distribution and view them as selection scores.
4
Experiments
In this section we describe our experiments in detail.
4.1
Data set
Although an active learning system doesn’t require
all the unannotated instances to be labeled initially,
having such an annotated data set is very useful
for simulations since it allows us to conduct experiments to inverstigate active learning, in our case,
for judging the quality of automatically generated
questions. To this end, we used the existing data
set called Mind the Gap data set which was created
and made publicly available by Becker et al. (2012)
3 . The data set consists of 2,252 questions generated
using sentences extracted from 105 Wikipedia’s articles across historical, social, and scientific topics.
Each question was rated by four Amazon Mechanical Turkers as Good, Okay, or Bad (see definitions
in Section 1.1).
For our experiments, we binarized the questions
into positive and negative examples. We considered
a question positive when all of its ratings were Good
or at most one rating was Okay or Bad. The rest of
the questions were considered as negative examples.
This way we obtained 747 positive and 1,505 were
negative examples. The chosen requirement for being a positive example was needed in order to focus
on high quality questions.
4.2
Features
In order to build models to judge the quality of questions, we implemented five types of features as in
Becker et al. (2012) including Token Count, Lexical, Syntatic, Semantic and Named Entity. In total
we had 174 features which are summarized in Table
2. The numbers inside parentheses are the indices of
the features used.
Questions with many gaps (with many missing
words) are harder to answer. Similarly, gaps with
many overlapped words with the remaining words
in the question are not suitable since they can be
easily inferred from the context. We used 5 different Token Count features to capture such properties.
We also used 9 Lexical features to capture different
3
http://research.microsoft.com/˜sumitb/questiongeneration
Type
Token Count - 5
Features
no. of tokens in answer(1) and in sentence(2), % of tokens in answer (3), no.(4)
and %(5) of tokens in answer matching with non-answer tokens
Lexical - 9
% of tokens in answer that are capitalized words(6), pronouns(7), stopwords(8),
and quantifiers(9), % of capitalized words(10) and pronouns(11) in sentence that
are in answer, does sentence start with discourse connectives ?(12), does answer
start with quantifier ?(13), does answer end with quantifier ?(14)
Syntatic - 116
is answer before head verb ? (15), depth of answer span in constituent parse
tree (16), presence/absence of POS tags right before the answer span(17-54),
presence/absence of POS tags right after the answer span(55-92), no. of tokens
with each POS tag in answers(93-130)
Semantic - 34
Answer covered by (131-147), answer contains(148-164) the semantic roles: {A0,
A1, A2, A3, A4, AM-ADV, AM-CAU, AM-DIR, AM-DIS, AM-LOC, AM-MNR,
AM-PNC, AM-REC, AM-TMP, CA0, CA1, Predicate}
Named Entities - 11
does answer contain a LOC(165), PERS(166), and ORG(167) named entities ? does
non-answer span contain a LOC(168), PERS(169), and ORG(164) named entities ?
no. (170) and % (171) of tokens in answer that are named entities, no. (172) and
% (173) of tokens in sentence that are named entities, % of named entities
in sentence present in answer (174)
Table 2: List of features used
statistics of pronouns, stop words, quantifiers, capitalized words, and discourse connectives. Similarly,
we used 116 Syntatic features that include mostly
binary features indicating presence/absence of a particular POS tag just before the gap and just after the
gap, and number of occurrences of each POS tag
inside the gap. Our semantic features includes 34
binary features indicating whether the answer contained a list of semantic roles and whether semantic roles cover the answer. In addition, we used 11
Named Entities features to capture presence/absence
of LOC, PERS and ORG entities inside the answer
and outside the answer. We also computed the entity density i.e. number of named entities present in
the answer. We used Senna tool for getting semantic
roles (Collobert et al., 2011) and Stanford CoreNLP
package (Manning et al., 2014) for getting POS tags
and named entities.
5
Results and Discussions
We conducted a number of experiments to see how
active learning performs at judging the quality of
201
Figure 2: Full Simulation for Na¨ıve Bayes Accuracy
questions at different settings: type of classifiers
(simple and committee), evaluation metrics (accuracy and F-Measure), seed data size, batch size, and
sampling algorithms. An experiment consists of a
number of runs. In each run, we divided the data
set into three folds using stratified sampling. We
considered one of the folds as the test data set and
Figure 3: Close-up view of Na¨ıve Bayes Accuracy
Figure 4: Full Simulation for Na¨ıve Bayes F1
Figure 5: Close-up view of Na¨ıve Bayes F1
merged the other two to construct the unlabeled data
set (U ). Remember that our data set is already labeled but we pretended that it is unlabeled U . Typically, the selected instances from U have to be labeled by a human. Since we already know all the
labels in the data set, we mimic the human labeling
202
by simply using the existing labels. This allows us
to conduct several experiments very efficiently.
In the first experiment, we compared the various
sampling techniques in terms of their impact of the
overall performance of question quality classifier.
To this end, we randomly selected 8 examples (four
positive and 4 negative) from U for the seed data set,
removed them from U and put them into the labeled
data set (L). We then built a Na¨ıve Bayes model for
judging the quality of questions using L. All the machine learning algorithms we used are available in
Weka (Hall et al., 2009). Next, we applied a given
sampling strategy to select 4 best examples (i.e. a
batch of size 4) to be labeled. These new labeled
examples were added to L and the question quality
classifier was retrained with this extended data set.
We used the test data subset to evaluate the question
quality classifier at each iteration and report accuracy and F-measure. The process was iterated until
the unlabeled data set U was empty.
We used the four sampling algorithms (i.e. least
confidence, minimum margin, entropy and random)
and report results in terms of average across 100 different runs; in each such run we ran the active learning approach entirely on all the data we had available. Figure 2 and Figure 4 present the accuracy and
F1 scores of Na¨ıve Bayes for each of the sampling
algorithms with respect to the number of labeled instances used. Figure 3 and Figure 5 are close-ups of
leftmost part of the curves in Figure 2 and Figure 4,
respectively. As we can see, all uncertainty sampling
methods (Min-margin, Entropy and Least confident)
outperformed random sampling for both accuracy
and F1 measures after few annotations were made.
For instance, with 200 examples selected by active
learning, the model provided 10% more in accuracy
and 4% more in F1 measure compared to the case
when the same number of instances were used by
sample randomly. It is a promising observation that
can save annotation budgets significantly. Moreover,
close-up graphs show that all three uncertainty sampling approaches rival each other. Note that all the
sampling methods converged (i.e. have same accuracy and F1 measure) at the end of the simulation. It
is normal because they would have the same set of
labeled instances by then.
In the second experiment, we formed a committee
of three probabilistic classifiers provided by Weka:
Na¨ıve Bayes, Logistic Regression, and SMO. These
classifiers learnt different hypotheses from the same
set of training examples. As discussed in Section
3.1.1, we generated three models from the same labeled set of examples and computed mean probability distributions. For this experiment, we set seed
size of 8, batch size of 4, and 100 runs as in experiment 1 and measured the performances of the
sampling algorithms. Figure 6 and Figure 8 show
the accuracy and F-measure for several sampling
strategies as a function of the number of annotated
examples. Figure 7 and Figure 9 are the close-up
views for Figure 6 and Figure 8 respectively. Again,
the uncertainty based sampling algorithms are very
competitive to each other and they outperform random sampling significantly in both accuracy and Fmeasure. This suggests that committee based active
learning is also useful for checking question quality.
To get an idea of the level of annotation savings when using active learning, consider we have
a budget for annotating about 160 instances. With
this budget (in Figure 6), uncertainty sampling algorithms provide 70% accuracy whereas random sampling provides only 65% accuracy. To attain 70%
accuracy, random sampling needs at least 360 samples (i.e. 200 examples more) to be labeled. With
360 samples, uncertainty sampling algorithms provide 74% accuracy. Similar observations can be
made when focusing on the F-measure. These observations clearly show the effectiveness of using active learning for judging the quality of automatically
generated questions.
Figure 7: Close-up view of Committee Accuracy
Figure 8: Full Simulation for Committee F1
Figure 9: Close-up view of Committee F1
Figure 6: Full Simulation for Committee Accuracy
In a third experiment, we focused on the effect of
203
the batch size on the behavior of the active learning approach. Note that we generate a new model
as soon as a new batch of labeled instances is ready.
For instance, a batch size of 2 means as soon as the
annotators provide two annotated instances, we add
them to the labeled set and generate a new model
from all the available labeled instances. The new
model is generally a better one as it is trained on a
larger training set than the previous one. However,
the smaller the batch size the larger the computational cost because we need to generate a model frequently. So, a balance between the computation cost
and the better model should be determined.
In the last experiment, we varied initial seed size
to see its effect of the initial seed size on our active learning approach. We experimented with seed
sizes of 4, 8, 16 and 32. We applied Na¨ıve based active learning with the batch size of 4 and 100 runs.
The plot in Figure 11 shows F1 measures of Entropy
based sampling at different seed set sizes. It can be
seen that the smaller the seed size, the smaller the
F1 score initially. Having a larger seed data initially
is beneficial which is obvious because in general the
larger the training set the better. We also included
a plot of the F1 measure corresponding to random
sampling with 32 seeds in Figure 11. It is interesting to note that although random sampling with 32
seeds has larger F1 score initially, it eventually performs poorly when more data is added.
6
Figure 10: Effect of Batch Size
To this end, we chose Na¨ıve based active learning
with entropy based sampling. We varied the batch
size from 1, 2, 4 and 8 and ran the experiment for
50 runs. The plot can be seen in Figure 10. As the
plot suggests, the performances are less sensitive to
batch sizes. A reasonable choice could be a batch
size of 4. But again, it depends on the amount of
computation cost available for model construction.
Conclusion
In this paper, we proposed to use active learning for
training classifiers for judging the quality of automatically generated gap-fill questions, which is the
first attempt of its kind to the best of our knowledge.
Our experiments showed that active learning is very
useful for creating cost-efficient methods for training question quality classifiers. For instance, it is observed that a reasonably good classifier can be built
with 300-500 labeled examples using active learning
(a potential stopping criteria) that can provide about
5-10% more in accuracy and about 3-5% more in
F1-measure than with random sampling. Indeed, the
proposed approach can accelerate the question generation process, saving annotation time and budget.
Although the proposed method is investigated in
the context of judging the quality of gap-fill questions, the method is general and can be applied
to other types of questions e.g., stem generation
for multiple choice questions and ranking of freeresponse questions. We plan implement the remaining steps (i.e. sentence selection and candidate
generation) of the question generation pipeline and
make it a complete system.
Acknowledgments
This research was partially sponsored by The University of Memphis and the Institute for Education
Sciences (IES) under award R305A100875 to Dr.
Vasile Rus.
Figure 11: Effect of seed data
204
References
Lee Becker, Sumit Basu, and Lucy Vanderwende. 2012.
Mind the gap: learning to choose gaps for question
generation. In Proceedings of the 2012 Conference
of the North American Chapter of the Association for
Computational Linguistics: Human Language Technologies, pages 742–751. Association for Computational Linguistics.
Jonathan C Brown, Gwen A Frishkoff, and Maxine Eskenazi. 2005. Automatic question generation for vocabulary assessment. In Proceedings of the conference
on Human Language Technology and Empirical Methods in Natural Language Processing, pages 819–826.
Association for Computational Linguistics.
Jinying Chen, Andrew Schein, Lyle Ungar, and Martha
Palmer. 2006. An empirical study of the behavior
of active learning for word sense disambiguation. In
Proceedings of the main conference on Human Language Technology Conference of the North American
Chapter of the Association of Computational Linguistics, pages 120–127. Association for Computational
Linguistics.
Ronan Collobert, Jason Weston, L´eon Bottou, Michael
Karlen, Koray Kavukcuoglu, and Pavel Kuksa. 2011.
Natural language processing (almost) from scratch.
The Journal of Machine Learning Research, 12:2493–
2537.
Sarah H Eason, Lindsay F Goldberg, Katherine M Young,
Megan C Geist, and Laurie E Cutting. 2012. Reader–
text interactions: How differential text and question
types influence cognitive skills needed for reading
comprehension. Journal of educational psychology,
104(3):515.
Arthur C Graesser, Shulan Lu, George Tanner Jackson,
Heather Hite Mitchell, Mathew Ventura, Andrew Olney, and Max M Louwerse. 2004. Autotutor: A tutor
with dialogue in natural language. Behavior Research
Methods, Instruments, & Computers, 36(2):180–192.
Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard
Pfahringer, Peter Reutemann, and Ian H Witten. 2009.
The weka data mining software: an update. ACM
SIGKDD explorations newsletter, 11(1):10–18.
Michael Heilman and Noah A Smith. 2009. Question generation via overgenerating transformations and
ranking. Technical report, DTIC Document.
Michael Heilman and Noah A Smith. 2010. Good
question! statistical ranking for question generation.
In Human Language Technologies: The 2010 Annual
Conference of the North American Chapter of the Association for Computational Linguistics, pages 609–
617. Association for Computational Linguistics.
Ayako Hoshino and Hiroshi Nakagawa. 2005. A realtime multiple-choice question generation for language
205
testing: a preliminary study. In Proceedings of the second workshop on Building Educational Applications
Using NLP, pages 17–20. Association for Computational Linguistics.
Ming Liu, Rafael A Calvo, and Vasile Rus. 2012. Gasks: An intelligent automatic question generation
system for academic writing support. Dialogue & Discourse, 3(2):101–124.
Christopher D Manning, Mihai Surdeanu, John Bauer,
Jenny Finkel, Steven J Bethard, and David McClosky.
2014. The stanford corenlp natural language processing toolkit. In Proceedings of 52nd Annual Meeting of
the Association for Computational Linguistics: System
Demonstrations, pages 55–60.
Karen Mazidi and Rodney D Nielsen. 2014a. Linguistic considerations in automatic question generation.
In Proceedings of Association for Computational Linguistics, pages 321–326.
Karen Mazidi and Rodney D Nielsen. 2014b. Pedagogical evaluation of automatically generated questions. In
Intelligent Tutoring Systems, pages 294–299. Springer.
Rada Mihalcea. 2004. Graph-based ranking algorithms
for sentence extraction, applied to text summarization.
In Proceedings of the ACL 2004 on Interactive poster
and demonstration sessions, page 20. Association for
Computational Linguistics.
Ruslan Mitkov, Le An Ha, and Nikiforos Karamanis.
2006. A computer-aided environment for generating
multiple-choice test items. Natural Language Engineering, 12(2):177–194.
Nobal B Niraula, Vasile Rus, Dan Stefanescu, and
Arthur C Graesser. 2014. Mining gap-fill questions
from tutorial dialogues. pages 265–268.
MJ Pollock, CD Whittington, and GF Doughty. 2000.
Evaluating the costs and benefits of changing to caa.
In Proceedings of the 4th CAA Conference.
Eric Ringger, Peter McClanahan, Robbie Haertel, George
Busby, Marc Carmen, James Carroll, Kevin Seppi, and
Deryle Lonsdale. 2007. Active learning for part-ofspeech tagging: Accelerating corpus annotation. In
Proceedings of the Linguistic Annotation Workshop,
pages 101–108. Association for Computational Linguistics.
Vasile Rus and Arthur C Graesser. 2009. The question
generation shared task and evaluation challenge. In
The University of Memphis. National Science Foundation.
Vasile Rus, Brendan Wyse, Paul Piwek, Mihai Lintean, Svetlana Stoyanchev, and Cristian Moldovan.
2010. The first question generation shared task evaluation challenge. In Proceedings of the 6th International Natural Language Generation Conference,
pages 251–257. Association for Computational Linguistics.
Vasile Rus, Nobal Niraula, Mihai Lintean, Rajendra Banjade, Dan Stefanescu, and William Baggett.
2013. Recommendations for the generalized intelligent framework for tutoring based on the development
of the deeptutor tutoring service. In AIED 2013 Workshops Proceedings, volume 7, page 116.
Burr Settles and Mark Craven. 2008. An analysis of active learning strategies for sequence labeling tasks. In
Proceedings of the Conference on Empirical Methods
in Natural Language Processing, pages 1070–1079.
Association for Computational Linguistics.
Dan Shen, Jie Zhang, Jian Su, Guodong Zhou, and ChewLim Tan. 2004. Multi-criteria-based active learning
for named entity recognition. In Proceedings of the
42nd Annual Meeting on Association for Computational Linguistics, page 589. Association for Computational Linguistics.
Jason D Williams, Nobal B Niraula, Pradeep Dasigi,
Aparna Lakshmiratan, Carlos Garcia Jurado Suarez,
Mouni Reddy, and Geoff Zweig. 2015. Rapidly scaling dialog systems with interactive learning.
206